Slideshare uses cookies to improve functionality and performance, and to provide you with relevant advertising. If you continue browsing the site, you agree to the use of cookies on this website. See our User Agreement and Privacy Policy.

Slideshare uses cookies to improve functionality and performance, and to provide you with relevant advertising. If you continue browsing the site, you agree to the use of cookies on this website. See our Privacy Policy and User Agreement for details.

Successfully reported this slideshow.

No Downloads

Total views

492

On SlideShare

0

From Embeds

0

Number of Embeds

3

Shares

0

Downloads

25

Comments

0

Likes

1

No embeds

No notes for slide

- 1. ISSN: 2278 – 1323 International Journal of Advanced Research in Computer Engineering & Technology Volume 1, Issue 4, June 2012 A Comprehensive Review of Image Smoothing Techniques Aditya Goyal, Akhilesh Bijalwan, Mr. Kuntal Chowdhury will not work efficiently when the noise rate is above 0.5. Abstract— Smoothing is often used to reduce noise within an Yang and Toh [2] used heuristic rules for improving theimage or to produce a less pixelated image. Image smoothing is performance of traditional multilevel median filter. Russoa key technology of image enhancement, which can remove and Ramponi [3] applied heuristic knowledge to build fuzzynoise in images. So, it is a necessary functional module in rule based operators for smoothing, sharpening and edgevarious image-processing software. Excellent smoothing detection. They can perform smoothing efficiently but not inalgorithm can both remove various noises and preserve details.This paper analyzed some image smoothing algorithms. These brightness. Choi and Krishnapuram [4] used a powerfulalgorithms have the ability of preserving details, such as robust approach to image enhancement based on fuzzy logicgradient weighting filtering, self-adaptive median filtering, approach, which can remove impulse noise, smoothing outrobust smoothing and edge preserving filtering. Appropriate non-impulse noise, and preserve edge well.choice of such techniques is greatly influenced by the imaging Different factors can cause different kinds of noise. Inmodality, task at hand and viewing conditions. This paper will practice, an image usually contains some different types ofprovide an overview of underlying concepts, along with noise. So good image smoothing algorithm should be able toalgorithms commonly used for image smoothing. deal with different types of noise. However, image smoothing [6] often causes blur and offsets of the Index Terms— Bilateral Filtering, Guided Filtering, Image edges. While the edge information is much important forsmoothing, Salt and Pepper Noise, Robust. image analysis and interpretation. So, it should be considered to keep the precision of edge‟s position in image smoothing. I. INTRODUCTION II.THE ANALYSIS OF SOME SMOOTHINGThe main aim of image smoothing is to remove noise in ALGORITHMSdigital images. It is a classical matter in digital imageprocessing to smooth image. And it has been widely used in A. Gaussian Smoothingmany fields, such as image display, image transmission andimage analysis, etc. Image smoothing [5] has been a basic The Gaussian smoothing operator is a 2-D convolutionmodule in almost all the image processing software. So, it is operator that is used to `blur images and remove detail andworth studying more deeply. noise. The idea of Gaussian smoothing is to use this 2-DImage smoothing is a method of improving the quality of distribution as a `point-spread function, and this is achievedimages. The image quality is an important factor for the by convolution. Since the image is stored as a collection ofhuman vision point of view. The image usually has noise discrete pixels we need to produce a discrete approximationwhich is not easily eliminated in image processing. The to the Gaussian function before we can perform thequality of the image is affected by the presence of noise. convolution. In theory, the Gaussian distribution is non-zeroMany methods are there for removing noise from images. everywhere, which would require an infinitely largeMany image processing algorithms can‟t work well in convolution kernel, but in practice it is effectively zero morenoisy environment, so image filter is adopted as a than about three standard deviations from the mean, and sopreprocessing module. However; the capability of we can truncate the kernel at this point. Figure 3 shows aconventional filters based on pure numerical computation suitable integer-valued convolution kernel that approximates a Gaussian with a of 1.0.is broken down rapidly when they are put in heavily noisy Once a suitable kernel has been calculated, then the Gaussianenvironment. Median filter is the most used method [1], but it smoothing can be performed using standard convolution methods. The convolution can in fact be performed fairly Manuscript received June, 2012. quickly since the equation for the 2-D isotropic Gaussian Aditya Goyal, Computer Science and Engineering, Uttarakhand shown above is separable into x and y components. Thus theTechnical University/ Dehradun Institute of Technology / Dehradun, India, / 2-D convolution can be performed by first convolving with a9997030797, (e-mail:goyaladit@gamil.com). 1-D Gaussian in the x direction, and then convolving with Akhilesh Bijalwan, Computer Science and Engineering, UttarakhandTechnical University/ Dehradun Institute of Technology / Dehradun, India, / another 1-D Gaussian in the y direction. The Gaussian is in8899806686, (e-mail:bijalwanakhilesh@gmail.com). fact the only completely circularly symmetric operator which Mr. Kuntal Chowdhury, Computer Science and Engineering, can be decomposed in such a way. The y component isUttarakhand Technical University/ Dehradun Institute of Technology / exactly the same but is oriented vertically.Dehradun, India, /8650198209 (e-mail: ikuntal09@gmail.com). 315 All Rights Reserved © 2012 IJARCET
- 2. ISSN: 2278 – 1323 International Journal of Advanced Research in Computer Engineering & Technology Volume 1, Issue 3, May 2012 when a pixel (often on an edge) has few similar pixels around it, the Gaussian weighted average is unstable.Where x, y are the spatial dimensions. Another issue concerning the bilateral ﬁlter is its eﬃciency. The brute-force implementation is in O(N r2 ) time, which is prohibitively high when the kernel radius r is large. In [14] an approximated solution is obtained in a discretized space-color grid. Recently, O(N ) time algorithms [8 , 9] have been developed based on histograms. Adams et al. [10] propose a fast algorithm for color images. All the above methods require a high quantization degree to achieve satisfactory speed, but at the expense of quality degradation.Original Image sigma = 20 sigma = 40 Figure 1. Example of Gaussian smoothingB. Edge Preserved FilteringTypical edge preserved filtering includes two types:Kuwahara filtering and selective mask filtering. The basicprocess of them is described as follows: Firstly, somedifferent templates are made based on the center pixel.Secondly, the mean value and the standard deviation of thepixels in different templates are calculated. Finally, the grayvalue of the center pixel is defined as the mean value in thetemplate where the standard deviation is the least. TheKuwahara filtering selects one template from 4 squarewindows. While the selective mask filtering chooses onetemplate region from 9 windows, which include 4 pentagon,4 hexagon and 1 square .The effect of preserving details ofthe selective mask filtering is better than that of the Figure 2. Example of Bilateral FilteringKuwahara filtering, because the former has more finewindow choice. Through the analysis of these algorithms, D. Optimization-Based Image Filteringwe find that the image details can be preserved fromselecting a suitable template according to the rule of A series of approaches optimize a quadratic cost function andminimizing the standard deviation. When the template is solve a linear system, which is equivalent to implicitlyselected, we can use other smoothing algorithms with ﬁltering an image by an inverse matrix. In imagebetter smoothing effects. For example, considering that segmentation and colorization , the aﬃnities of this matrixthe medium filtering can preserve details more effectively are Gaussian functions of the color similarities. In imagethan mean filter does and that it is more effective for salt and matting, a matting Laplacian matrix is designed to enforcepepper noise, we can take the medium grey value in the the alpha matte as a local linear transform of the imagetemplate window where the standard deviation is the least colors. This matrix is also applicable in haze removal [12].as the grey value on the centre pixel. And so we can remove The weighted least squares (WLS) ﬁlter in [11] adjusts thesalt and pepper noise more effectively. matrix aﬃnities according to the image gradients and produces a halo-free decomposition of the input image.C. Bilateral Filter Although these optimization-based approaches often generate high quality results, solving the correspondingThe bilateral ﬁlter [7] computes the ﬁlter output at a pixel as a linear system is time-consuming.weighted average of neighboring pixels. It smoothes the It has been found that the optimization-based ﬁlters areimage while preserving edges. Due to this nice property, it closely related to the explicit ﬁlters. In [15] Elad shows thathas been widely used in noise reduction , HDR compression , the bilateral ﬁlter is one Jacobi iteration in solving themulti-scale detail decomposition , and image abstraction . It Gaussian aﬃnity matrix. In [16] Fattal deﬁnes theis gen- eralized to the joint bilateral ﬁlter in , in which the edge-avoiding wavelets to approximate the WLS ﬁlter. Theseweights are computed from another guidance image rather explicit ﬁlters are often simpler and faster than thethan the ﬁlter input. The joint bilateral ﬁlter is particular optimization-based ﬁlters.favored when the ﬁlter input is not reliable to provide edge E. Nonlinear Diffusion Filteringinformation, e.g., when it is very noisy or is an intermediateresult. The joint bi- lateral ﬁlter is applicable in ﬂash/no-ﬂash NLDF was originally formulated by Perona and MalikDenoising, image upsamling , and image deconvolution . (1987). Noise is smoothed locally, ‟within‟regions definedHowever, it has been noticed [11, 13, 14] that the bilateral by object boundaries whereas little or no smoothing occursﬁlter may have the gradient reversal artifacts in detail between image objects. Local edges are enhanced sincedecomposition and HDR compression. The reason is that discontinuities, such as boundaries, are amplified. 316 All Rights Reserved © 2012 IJARCET
- 3. ISSN: 2278 – 1323 International Journal of Advanced Research in Computer Engineering & Technology Volume 1, Issue 4, June 2012Mathematically one treats the problem like a diffusionprocess, where the diffusion coefficient is adapted locally tothe effect that diffusion stops as soon as an object boundary isreachedDiffusion can be thought of as the physical process thatequilibrates concentration differences without creating ordestroying mass. Mathematically, this is described by Fick‟slaw:where the flux j is generated to compensate for theconcentration gradient u.D is a tensor that describes therelation between them. Now, using the Continuity Equation(Conservation of mass):We get:The solution of the linear diffusion equation with a scalardiffusivity d i s exactly the same operation as convolving the Figure 4. Comparison of low pass filtering to non-diffusion filtering.image u with a Gaussian kernel of width . F. Robust Smoothing Filter Robust smoothing filter is a simple and fast nonlinear filter. ItPerona and Malik proposed to exchange the scalar diffusion can remove salt and pepper noise with lower densityconstant d with a scalar-valued function g of the gradient u effectively. Because it adopts the strategy of losing the abilityof the grey levels in the image. The diffusion equation then of smoothing to preserve edges, it can preserve more edgereads: details than the medium filtering can do. The process can be described as follows: （1）Calculate the maximum and the minimum of the gray values in the template window except the centre pixel. （2）Compare the gray value on the centre pixel with theThe length of the gradient is a good measure of the edgestrength of the current location which is dependent on the maximum and the minimum.differential structure of the image. This dependence makes （3）If the gray value is larger than the maximum , thethe diffusion process nonlinear. maximum is output; If the gray value is smaller than the minimum, the minimum is output; If the gray value is between them, the gray value is output. In order to preserve more details, we adopt the same idea in robust smoothing filtering as in the adaptive medium filtering. It can reduce the image distortion by unchanging the gray onv„middle level‟ pixels.Original image smoothed image Figure 3. Example of non-diffusion Filtering （a) the image corrupted with salt and pepper noise 317 All Rights Reserved © 2012 IJARCET
- 4. ISSN: 2278 – 1323 International Journal of Advanced Research in Computer Engineering & Technology Volume 1, Issue 3, May 2012 generic concept and is applicable in other applications beyond the scope of ”smoothing”. Moreover, the guided ﬁlter has an O(N) time (in the number of pixels N) exact algorithm for both gray-scale and color images. Experiments show that the guided ﬁlter performs very well in terms of both quality and eﬃciency in a great variety of applications, such as noise reduction, detail smoothing/enhancement, HDR compression, image matting/feathering, haze removal, and joint up sampling. 1. Guided Filter Kernel(b) The result smoothed by robust smoothing method Figure 5. Example of Robust Smoothing method We first define a general linear translation- variant filteringG. Gradient Weighting Filter process, which involves a guidance image I, an input image p, and output image q. Both I and p are given beforehandThe gradient-dependent weighting filters are mainly based on according to the application, and they can be identical. Thethe following principle: in a discrete image, the difference of filtering output at a pixel I is expressed as a weightedthe gray values on pixels in outer area is larger than that in average:-inner area. In same area, the change on centre pixels issmaller than that on edge pixels. The gray gradient is direct qi= ∑j Wij (I) pj. (4)ratio to the gray difference in vicinity. That is, where the graychange is slower, the gradient is smaller. A function whose where i and j are pixel indexes. The filter kernel Wij is avalue reduces with the increase of the gradient is adopted, function of the guidance image I and independent of p. Thisand it is chosen as the weight of the window. So, the filter is linear with respect to p.smoothing contribution is mainly coming from the same area. The guided filtering kernel Wij is given by:-Accordingly the edge and the detail cannot be lost apparentlyafter image smoothing.In designing gradient-selected filters, power and exponential Wij(I)= ) (5)function are often chosen as weighting function. Especiallywhen the power is equal to –1, the filtering is called gradientreciprocal weighting filtering. When the function is the Where I is guidance image, p is input image, q is outputexponential one, the filtering is called adaptive filtering. image, Wij is filter kernel, is variance, Ki is normalizingWhen we extract lines from remote sensing images, the parameter, Wk is window centered at pixel k and is meanadaptive of I.filtering is often adopted in preprocessing to realize the aimof noise removal and edge enhancement. It can be describedas:Where, x is the gradient, k is the parameter that determinesthe smoothing degree.By analysis, we find that k can be used to adjust the degree ofsharp of the exponential function. If k is bigger, theexponential function will be slower in change. So, if thegradient is bigger than k, the gradient will increase withthe adding of the iterative times, so as to realize the aim ofsharping edge. Oppositely, if the gradient is smaller than k,the details will be smoothed. Thus, the value of k is critical tothe smoothing effect.H. Guided Image FilterGuided image filter [7] is an explicit image filter, derivedfrom a local linear model; it generates the ﬁltering output byconsidering the content of a guidance image, which can bethe input image itself or another diﬀerent image. Moreover,the guided ﬁlter has a fast and non-approximate linear-timealgorithm, whose computational complexity is independentof the ﬁltering kernel size. The guided filter output is locally a Figure 6. Guided image filter against other filters.linear transform of the guidance image. This ﬁlter has theedge-preserving smoothing property like the bilateral ﬁlter, III. CONCLUDING REMARKSbut does not suﬀer from the gradient reversal artifacts. It isalso related to the matting Laplacian matrix, so is a more 318 All Rights Reserved © 2012 IJARCET
- 5. ISSN: 2278 – 1323 International Journal of Advanced Research in Computer Engineering & Technology Volume 1, Issue 4, June 2012Image smoothing algorithms offer a wide variety ofapproaches for modifying images to achieve visuallyacceptable images. The choice of such techniques is afunction of the specific task, image content, observercharacteristics, and viewing conditions. There are two majorshortcomings of Gaussian smoothing, it shrinks shapes anddislocates boundaries when moving from finer to coarserscales, blurs important image features. Bilateral filterproduces staircase effect, bilateral filter tends to removetexture, create flat intensity regions and new contours. Therobust smoothing filtering only can remove salt and peppernoise with lower density. Nonlinear diffusion filteringapproach smooth‟s regions of low brightness gradient whileregions of high gradients are not smoothed. These anisotropicdiffusion approaches are contrast-driven and smooth globallysalient but low contrast image features. The guided ﬁlter hasa fast and non-approximate linear-time algorithm, whosecomputational complexity is independent of the ﬁlteringkernel size. As a locally based operator, the guided ﬁlter isnot directly applicable for sparse inputs like strokes. It alsoshares a common limitation of other explicit ﬁlter - it mayhave halos near some edges.Although we did not discuss the computational cost of imagesmoothing algorithms in this article it may play a critical rolein choosing an algorithm for real-time applications. Despitethe effectiveness of each of these methods when appliedseparately, in practice one has to devise a combination ofsuch methods to achieve more effective image smoothing.We believe that the simplicity and eﬃciency of the guidedﬁlter still make it beneﬁcial for image smoothing. REFERENCES[1] K. Arakawa, “Median filter based on fuzzy rules and its application toimage restoration,” Fuzzy Sets And Systems, vol. 77, (1996)pp. 3-13. Aditya Goyal, B.Tech (IT), MBA (Marketing) and pursuing M.Tech[2] X. Yang and P-S Toh, “Adaptive fuzzy multilevel filter,” IEEE Trans. (CSE) from Dehradun Institute of Technology Dehradun.onImage Processing,” vol. 4, no. 5, (1995)pp. 680-682.[3] F. Russo and G. Ramponi, “A fuzzy operator for the enhancement of Akhilesh Bijalwan, M.Sc (IT) and pursuing M.Tech (CSE) fromblurred And noisy image,” IEEE Trans. on Image Processing, vol. 4, no. 8, Dehradun Institute of Technology Dehradun.(1995), pp. 1169-1174.[4] Y. S. Choi and R. Krishnapuram, “A robust approach to imageenhancement based on fuzzy logic,” IEEE Trans. on Image Processing, vol. Mr. Kuntal Chowdhury, Assistant Professor, Dehradun Institute of6, no. 6,(1997), pp. 808-825. Technology, Dehradun.[5] Qin Zhiyuan , Zhang Weiqiang , Zhang Zhanmu , Wu Bing , Rui Jie ,ZhuBaoshan “A robust adaptive image smoothing algorithm”.[6] Dillip Ranjan Nayak, “Image smoothing using fuzzy morphology,”International journal of computer application issue2, volume 1 (2012)[7] Hek., sun j., tang x.: Guided image filtering. In Proc. of the EuropeanConference on Computer Vision (2010), vol. 1, pp. 1–14.[8] Porikli, F.: Constant time o(1) bilateral ﬁ ltering. CVPR (2008)[9] Yang, Q., Tan, K.H., Ahuja, N.: Real-time o(1) bilateral ﬁ ltering. CVPR(2009)[10] Adams, A., Gelfand, N., Dolson, J., Levoy, M.: Gaussian kd-trees forfast high-dimensional filtering. SIGGRAPH (2009)[11] Farbman, Z., Fattal, R., Lischinski, D., Szeliski, R.: Edge-preservingdecomposi-tions for multi-scale tone and detail manipulation. SIGGRAPH(2008)[12] He, K., Sun, J., Tang, X.: Single image haze removal using dark channelprior.CVPR (2009)[13] Durand, F., Dorsey, J.: Fast bilateral ﬁ ltering for the display ofhigh-dynamic-rangeimages. SIGGRAPH (2002)[14] Bae, S., Paris, S., Durand, F.: Two-scale tone management forphotographic look.SIGGRAPH (2006)[15] Elad, M.: On the origin of the bilateral ﬁ lter and ways to improve it.IEEE Transactions on Image Processing (2002)[16] Fattal, R.: Edge-avoiding wavelets and their applications. SIGGRAPH(2009). 319 All Rights Reserved © 2012 IJARCET

No public clipboards found for this slide

×
### Save the most important slides with Clipping

Clipping is a handy way to collect and organize the most important slides from a presentation. You can keep your great finds in clipboards organized around topics.

Be the first to comment